Tag
8 postsLLM Applications
Posts about LLM Applications by Kousigan A, filed under AI Coding. Each one links the official docs it relies on.
Deep dive13 minAI Coding
LLM feedback loop: turn production failures into eval cases
Run an LLM feedback loop: collect user feedback, read failing traces, name failure types, fix one layer, and keep each fixed failure as an eval case.
Deep dive10 minAI Coding
AI agent state management: resume a run without paying twice
Design state for an AI agent you build: task state in your own store, a checkpoint per step, tool calls safe to retry, and long-term memory that expires.
Deep dive14 minAI Coding
RAG retrieval: get the right chunk into the top k
Build RAG retrieval in steps: clean and chunk documents, filter by permission, use hybrid search, rerank, cite the chunk, and test retrieval on its own.
Deep dive10 minAI Coding
Valid JSON, wrong values: validate LLM structured outputs
Get reliable JSON from an LLM: structured outputs or strict tool calling, Pydantic validation, refusals and cut-offs, bounded retries and schema versions.
Deep dive12 minAI Coding
LLM model routing: test on evals, pin it, fall back safely
Choose an LLM per task with your own eval set, route by task type, classifier or failed check, keep fallbacks on the same contract, and pin versions.
Deep dive13 minAI Coding
Unit tests miss LLM output that varies: build an eval suite
Build an eval suite for an LLM feature: real and edge-case inputs, a grader per property, several runs per case, and a rerun on every prompt or model change.
Deep dive10 minAI Coding
LLM observability: trace every step, prompts off by default
Trace an LLM feature in production with OpenTelemetry: one trace per request, the fields to record, and the prompts and secrets to keep out.
Deep dive10 minAI Coding
Threat model an LLM feature before you give it tools
Threat model an LLM feature: map trust boundaries, run STRIDE, check the OWASP 2025 LLM risks, and keep authorisation in the tools and the data layer.